NormaFlood: Closed-Form Flood Forecasting Without Backpropagation

Abstract

This repository contains the code, data pipelines, and manuscript for NormaFlood, an algebraic streamflow forecasting engine evaluated on India's official CAMELS-IND benchmark (229 streamflow gauges, test period 2016–2020). Dispensing with iterative backpropagation and gradient descent, the model uses an exact regularized ridge solve combined with an online Recursive Least Squares (RLS) Woodbury update. The core architectural finding is that constructing autoregressive lag features in the Box-Cox transformed target space enables linear models to capture persistence and pulse dynamics, lifting median NSE from 0.519 (raw features) to 0.638 (Box-Cox space) and 0.658 (online adaptive). On the strict 163-basin long-record subset, performance reaches median NSE 0.709 (KGE 0.731). Calibrated exceedance probabilities for flood alarms (P(Q>Q95)P(Q>Q95)) are produced analytically from the parameter posterior covariance (Brier 0.014–0.016, AUC 0.970–0.973, Maximum Economic Value 0.844). All experiments (E1–E8) regenerate deterministically in seconds.